Physix Frontier badgePhysix Frontier
PublishChinese | EN

Physix Frontier · News Briefing Card (enbrief · Sep 22, 2026)

Huawei unveils Atlas 960 roadmap for trillion-parameter models

KEY FACTS

  • The Atlas 960 SuperPoD is planned to integrate 15,488 processors, delivering 30 EFLOPS of compute.
  • The Atlas 960E connects 4,096 NPUs via UnifiedBus, supporting training of trillion-parameter models.
  • It adopts near-package optics, cutting interconnect power consumption by more than 550 kilowatts.
  • The Ascend 960DT and 960PR chips are slated to launch in Q1 and Q3 of 2027, respectively.

KEY DATA

15,488Max processor count
30 EFLOPSFP8 compute
4,460 TBMemory capacity
34 PB/sInterconnect bandwidth

PHYSIX OBSERVATION

Huawei is shifting from single-point compute competition to system-level engineering breakthroughs. Facing process-node constraints, its strategy is to push cluster communication efficiency to the limit through the UnifiedBus architecture, trading scale for performance. This "stack hardware plus interconnect" approach can support trillion-parameter large models, but it poses severe challenges for data center cooling and energy management, marking that domestic AI infrastructure has entered deep waters.

Source: enbrief original report ↗